The AI landscape has undergone a dramatic transformation in the past year, fueled by the collaborative spirit of open source development. No longer solely the domain of tech giants, large language models (LLMs) are now evolving through community efforts and open sharing, impacting everything from infrastructure to algorithm optimization and deployment. This open source movement is accelerating AI’s progress, making it more accessible and democratizing the opportunity to contribute to the next generation of intelligent systems.
Against this backdrop, the GOSIM AI Paris 2025 conference, co-hosted by GOSIM, CSDN, and 1ms.ai, commenced on May 6th in Paris, France. The event serves as a crucial platform, connecting global technology practitioners and researchers to explore the latest breakthroughs and future directions in open source AI.
The conference boasts an impressive lineup of over 80 technology experts and scholars from leading organizations such as Alibaba, Hugging Face, BAAI, MiniMax, Neo4j, Dify, MetaGPT, Zhipu AI, Eigent.AI, Docker, Inflow, Peking University, Fraunhofer, Oxford University, and the French openLLM community. Key partners, including Huawei, the All-China Youth Innovation and Entrepreneurship Association in France, the Sino-French Artificial Intelligence Association, the Apache Software Foundation, the Eclipse Foundation, The Khronos Group, WasmEdgeRuntime, LF Generative AI Commons, the Linux Foundation Research, the OpenWallet Foundation, the Open Source Initiative (OSI), Software Heritage, and K8SUG, are also actively participating. The conference features over 60 technical sessions centered around core themes like AI models, infrastructure, application deployment, and embodied intelligence, providing a comprehensive view of the open source ecosystem’s evolution and emerging trends.
The Symbiotic Relationship Between AI and Open Source
Michael Yuan, co-founder of GOSIM, kicked off the conference with a keynote address titled “Open Source Has Caught Up, What’s Next?” He shared his insights on the current state and future trajectory of open source AI, emphasizing that it has reached a pivotal moment.
“We once predicted it would take 5-10 years for open source to catch up with closed-source models, but it seems that this goal has been achieved ahead of schedule,” Yuan stated. He cited the recent release of Qwen 3 as an example, noting that open source models are no longer just competing with each other but are now directly challenging proprietary flagship models, even surpassing them in certain benchmarks. Yuan also suggested that this progress is not solely due to open source advancements but also the result of closed-source development failing to meet expectations and encountering performance bottlenecks. In contrast, open source models are rapidly evolving, exhibiting a steep performance growth curve and demonstrating a true “catch-up” phenomenon.
This observation raises a fundamental question: How far are we from achieving Artificial General Intelligence (AGI)? Yuan believes that the future of AGI may not lie in a single, all-encompassing model but rather in a network of specialized models, knowledge bases, and tools deployed on private hardware or robotic devices.
He further elaborated that the AI architecture is shifting from a centralized to a decentralized paradigm. He highlighted OpenAI’s transition from the Completion API to the new Responses API, which aims to build a large-scale intelligent agent platform. Nearly 600,000 users and developers have already joined this transformation, contributing to the development of distributed AI applications.
“The future of AGI should not be exclusively developed by a single, well-funded company,” Yuan asserted. “Instead, it should be built through global collaboration, creating an ecosystem network encompassing models, knowledge bases, robots, and execution systems.”
Following Yuan’s address, Daniel Goldscheider, Executive Director of the OpenWallet Foundation, delivered a presentation on “GDC Wallets & Credentials,” focusing on the Global Digital Compact (GDC) project, adopted by the United Nations General Assembly. He explained that the GDC has two core objectives:
- Recognizing that digital technologies have profoundly transformed our lives and societal development, bringing both unprecedented opportunities and unforeseen risks.
- Emphasizing that realizing the full potential of digital technologies for the benefit of all humanity requires global cooperation, breaking down barriers between countries, industries, and even public and private sectors.
Based on this shared understanding, the GDC has spawned the “Global Digital Collaboration” initiative, aiming to foster genuine collaboration between governments, businesses, non-profit organizations, and other stakeholders.
When discussing the operational aspects, Goldscheider emphasized that this collaboration is not dictated by any single organization but rather adopts a “joint convening” approach, inviting all interested international organizations, standard-setting bodies, open source communities, and intergovernmental organizations to participate. He clarified that this is not a “who leads whom” project but an equal collaboration platform where every party has a voice and no one is more important than another.
He further explained that the Global Digital Collaboration does not aim to directly develop standards or technologies but rather to facilitate a dialogue among organizations from diverse backgrounds, allowing them to present their perspectives and needs to reach a consensus. Subsequently, the specific standards and technical work will be advanced by the relevant specialized bodies. He cited “digital identity” and “biometric technology” as examples, noting that many organizations are already working in these areas, highlighting the need for a neutral platform to bring everyone together, avoid duplication, conflicts, and resource wastage.
Four Dedicated Forums: A Comprehensive Analysis of Open Source AI
The conference featured four specialized forums: AI Models, AI Infrastructure, AI Applications, and Embodied Intelligence. These forums covered critical topics ranging from underlying architecture to application deployment, and from model capabilities to intelligent agent practices. Each forum hosted leading experts from global enterprises and research institutions, providing both in-depth analysis of the latest technological trends and showcasing rich engineering practice cases, demonstrating the comprehensive integration and evolution of open source AI across multiple fields.
Deconstructing the Underlying Logic of AI Large Models
The AI Models forum brought together experts from open source communities and research institutions to share insights on architectural innovations, open source collaboration, and ecosystem evolution in the realm of large models.
Guilherme Penedo, Machine Learning Research Engineer at Hugging Face, presented “Open-R1: A Fully Open Source Reproduction of DeepSeek-R1,” showcasing the Open-R1 project’s efforts in replicating the DeepSeek-R1 model, with a focus on promoting the openness and standardization of data related to inference tasks. Guang Liu, Technology Leader of the Data Research Team at Zhiyuan Research Institute, shared “OpenSeek: Collaborative Innovation Towards the Next Generation of Large Models,” emphasizing the importance of global collaboration in driving breakthroughs in model performance at the algorithm, data, and system levels, with the goal of developing the next generation of large models that surpass DeepSeek.
Jason Li, Senior Vice President of CSDN, delivered “Decoding DeepSeek: Technological Innovation and its Impact on the AI Ecosystem,” providing an in-depth analysis of DeepSeek’s innovations in technical paradigms, model architecture, and industrial ecology, as well as its potential impact on the global AI ecosystem. Yiran Zhong, Senior Research Director at MiniMax, presented “Linear Future: The Evolution of Large Language Model Architectures,” introducing the team’s proposed Lightning Attention mechanism, which offers a potential alternative to Transformer architectures in terms of efficiency and performance. Shiwei Liu, Royal Society Newton International Fellow at Oxford University, discussed “The Depth Curse in Large Language Models,” exploring the diminishing contributions of deep neural networks as models deepen, and proposing the use of LayerNorm Scaling to improve the Pre-LN mechanism to enhance deep layer utilization and overall efficiency. Diego Rojas, Research Engineer at Zhipu AI, pointed out in “Code Large Language Models: Exploring Beyond Tokens” that current large models, while powerful, still rely on tokenization, which is inefficient, and shared new methods for skipping tokenization to make models faster and stronger. Nicolas Flores-Herr, Head of the Basic Models Team at Fraunhofer IAIS, concluded the forum with “How to Build Globally Competitive ‘European-Made’ Large Language Models?” emphasizing that Europe is overcoming data, diversity, and regulatory challenges through multilingual, open source, and trustworthy localized large model projects, to build the next generation of AI that reflects European values.
The Triad of AI Infrastructure: Data, Computing Power, and Algorithmic Evolution
Focusing on building a more open, efficient, and inclusive foundation for large models, the AI Infrastructure forum brought together leading experts from research institutions and enterprises to engage in in-depth discussions on key issues such as data, computing power, and system architecture.
Yonghua Lin, Vice President of Zhiyuan Research Institute (BAAI), launched the Chinese Internet Corpus CCI 4.0 in “AI Open Source for Good: Inclusive Applications, Fair Data, and Universal Computing Power,” covering three major datasets: CCI4.0-M2-Base V1, CCI4.0-M2-CoT V1, and CCI4.0-M2-Extra V1. CCI4.0-M2-Base V1 has a data volume of 35000GB, is bilingual in Chinese and English, with 5000GB of Chinese data, a 5-fold increase in data scale compared to CCI3.0. CCI4.0-M2-CoT V1 contains 450 million reverse synthesized human thought trajectory data for improving reasoning ability, with a total token number of 425B (425 billion), nearly 20 times the size of Cosmopedia (open sourced by Hugging Face), the largest open source synthetic dataset currently available globally.
Xiyuan Wang, Senior Software Engineer at Huawei, then introduced how the CANN architecture connects AI frameworks and Ascend hardware in “Best Practices for Training and Inference Based on Ascend CANN,” and achieves optimal training inference through supporting ecosystems such as PyTorch and vLLM. Guillaume Blaquiere, Data Architect at Carrefour, demonstrated how to deploy serverless large model instances supporting GPUs through Google Cloud Run to reduce costs and improve resource utilization efficiency in “Making Your LLM Serverless.” Yinping Ma, Engineer at Peking University, gave a keynote speech on “Open Source Intelligent Computing Integrated Management and Scheduling Basic Software - SCOW and CraneSched,” introducing the two major open source basic software developed by Peking University, SCOW and CraneSched, which have been deployed in dozens of universities and enterprises across the country, supporting the unified management and high-performance scheduling of intelligent computing resources. Yaowei Zheng, PhD candidate at Beihang University, shared the design concept of the hybrid controller architecture in the Verl system in the “verl: A RLHF System Based on Hybrid Controller” speech, and discussed its efficiency advantages in large-scale reinforcement learning training. Greg Schoeninger, CEO of Oxen.ai, presented the “Training Datasets and Infrastructure for DeepSeek-R1 Style Reinforcement Learning (GRPO)” and detailed the practice path for reinforcement learning training processes for reasoning LLMs, including dataset construction, infrastructure building, and local training code generation models.
From “Can it be Used” to “Is it Used Well”: AI Applications Enter the Practical Stage
In the AI Applications forum, R&D practitioners and technology decision-makers from leading companies shared a diverse range of insights, showcasing the real-world deployment paths and future possibilities of AI applications driven by large models.
Yongbin Li, Chief Researcher at Alibaba Tongyi Lab, shared the latest progress of Tongyi Lingma in technical evolution and product application in “Tongyi Lingma: From Coding Copilot to Coding Agent.” Dongjie Chen, Software Engineer at Huawei, gave a keynote speech on “Cangjie Magic: A New Choice for Developers in the Era of Large Models,” introducing the AI large model Agent development framework based on the Cangjie programming language, which can significantly improve the efficiency of developers in building intelligent HarmonyOS applications and bring excellent development experience. Xinrui Liu, Director of the LangGenius Developer Ecosystem, focused on “Working Together, Technical Power Enabled by Dify,” emphasizing Dify’s open source ecosystem and its role in accelerating the popularization of AI applications.
Regarding the combination of AI and system engineering, Rik Arends, co-founder of Makepad, gave a unique presentation: “Using Ambient Coding, Use AI to Create Rust UI for Mobile Devices, Web Pages, and Mixed Reality,” exploring how to use ambient coding to build a new paradigm for UI. Christian Tzolov, R&D Software Engineer from the Broadcom Spring team, focused on demonstrating how to efficiently integrate AI models with existing systems and resources through the MCP Java SDK and Spring AI MCP in “A Unified Paradigm for AI Integration Through MCP.” Wenjing Chu, Senior Director of Technology Strategy at Futurewei, further elevated the perspective in “The ‘T’ in MCP and A2A Stands for Trust,” deeply analyzing how to build truly trustworthy AI systems in agent-based applications. In addition, Hong-Thai Nguyen, Software Engineering Manager at Cegid, introduced how multi-agent can reshape business processes and achieve smarter enterprise decision-making and operation in combination with practical scenarios in the “Cegid Pulse: Multi-Agent Business Management Platform” speech.
When Large Models are Equipped with “Bodies”: Embodied Intelligence Arrives
Embodied intelligence is becoming one of the most challenging and promising development directions in the field of AI. In this forum, many of the industry’s top technical experts engaged in in-depth discussions around the theme of “embodied intelligence,” sharing their practical explorations in architectural design, model application, and scenario deployment.
Angelo Corsaro, CEO and CTO of ZettaScale, introduced how the Zenoh protocol can break down the barriers between perception, execution, and cognition in the intelligent robot era in “Mind, Body, and Zenoh.” Philipp Oppermann, Project Manager of the Dora project, brought “Using Zenoh in Dora to Implement Distributed Data Flow,” explaining the important application of the Zenoh protocol in Dora to implement distributed data flow. James Yang, Professor at the University of Science and Technology of China, gave a speech on “Generation of Adversarial Safety-Critical Scenarios in Autonomous Driving,” introducing how to improve the safety of autonomous driving technology bygenerating adversarial scenarios to ensure stability and reliability in complex environments.
In addition, Minglan Lin, an embodied intelligence researcher at the Zhiyuan Research Institute, also focused on the topic of “RoboBrain: A Unified Brain Model for Robot Operation & RoboOS: A Hierarchical Collaboration Framework forRoboBrain and Robot Intelligent Agents,” demonstrating how RoboBrain can improve the intelligence level of robots and the important role of RoboOS in robot collaboration. Ville Kuosmanen, founder of Voyage Robotics, gave a wonderful speech on “Building Robot Applications with Open Source VLA Models,” explaining how to use open source VLA models to provide strong support for robot applications. Finally, Huy Hoang Ha, a large language model researcher at Menlo Research, discussed how spatial reasoning can help robots better understand complex 2D and 3D environments, thereby improving their operation and navigation capabilities in the keynote speech of “Spatial Reasoning LLM: Enhancing Understanding of 2D and 3D to Support Robot Operation and Navigation.”
Spotlight Talks: Illuminating Cutting-Edge Technologies and Innovative Applications
The Spotlight Talks Day 1 featured engaging presentations from industry experts on cutting-edge technologies and innovative applications. This segment served as a platform for technology practitioners from various domains to discuss the latest advancements and practical applications of AI. Cyril Moineau, Research Engineer at the French Atomic Energy Commission (CEA), introduced how the Eclipse Aidge project supports the deployment and optimization of deep neural networks on embedded platforms by providing a complete toolchain in the speech of “Aidge”, thereby accelerating the development of edge intelligent systems.
Paweł Kiszczak, Data Scientist at Bielik.ai, publicly shared the latest progress of the Polish native AI project Bielik for the first time at this conference, and gave a speech entitled “The Rise of Bielik.AI,” telling how the project promotes the construction of a local autonomous AI system through open source language models and a complete tool ecosystem. The Bielik project has not only released multiple open source language models (parameter scales covering 1.5B, 4.5B and 11B), but also created an end-to-end toolchain covering datasets, evaluation, training and fine-tuning, supporting research teams and developers to fine-tune or continuously pre-train based on basic models, which greatly reduces the R&D threshold for large models and stimulates local technology innovation capabilities.
Hung-Ying Tai, Technical Lead from Second State, shared “Running GenAI Models on Edge Devices with LlamaEdge,” demonstrating LlamaEdge’s lightweight and high-performance capabilities in deploying generative AI models on edge devices, bringing a more flexible and efficient local reasoning experience. Tianyu Chen, a PhD candidate at Peking University, introduced how the SAFE framework alleviates the problem of scarce training data through the self-evolution mechanism of “data synthesis-model fine-tuning,” thereby significantly improving the efficiency and accuracy of Rust code formal verification in “Achieving Automatic Formal Verification for Rust Code Based on Self-Evolution Framework.” Gautier Viaud, R&D Director at Illuin Technology, shared how the ColPali system, built by the team based on the ColBERT architecture and PaliGemma model, effectively improves the accuracy and efficiency of document retrieval by combining graphic and text information in the speech “ColPali: Efficient Document Retrieval Based on Visual Language Model.” Finally, Xiao Zhang, CEO of Dynamia.ai, introduced how to better manage and schedule heterogeneous GPU resources with the help of HAMi and improve the utilization rate and observability of AI infrastructure in “Unlocking the K8s Cluster Capabilities of Heterogeneous AI Infrastructure: Releasing the Power of HAMi.”
Diverse Interactions and Highlights of the First Day
In addition to the high-density keynote speeches, the conference also featured several special units. The Closed-door Meeting unit focused on strategic dialogues and in-depth industry exchanges to promote cross-border cooperation. The Showcase Sessions focused on presenting the latest AI technology products of enterprises and research institutions, attracting a large number of visitors to stop and communicate. In the Competition Sessions, AI and robotics developers, engineers, and robotics enthusiasts from around the world focused on the open source SO-ARM100 robotic arm kit to carry out practical exploration of imitation learning. The kit integrates Hugging Face’s LeRobot framework and combines NVIDIA’s AI and robotics technologies to support cutting-edge AI architectures including ACT and Diffusion Policy, providing participants with a solid technical foundation. Participants conducted practical explorations in real scenarios to comprehensively evaluate its effects and feasibility.
The Workshop Sessions took the OpenHarmony ecosystem as the core topic and explored the open source project incubated and operated by the Open Atom Open Source Foundation. OpenHarmony is committed to building an intelligent terminal operating system framework for the era of all-scenario, all-connection, and all-intelligence, creating an open, globalized, and innovative leading distributed operating system platform, serving diverse intelligent devices, and helping the development of the Internet of Everything industry. At the conference site, participants deeply understood the core advantages of OpenHarmony in multi-device collaboration and lightweight system design through a series of practical workshops, personally participating in key processes from driver development to application deployment. Hands-on practice not only helps developers open up the “bottom-to-end” technical path, but also comprehensively improves system-level development and debugging capabilities.
The GOSIM AI Paris 2025 Day 1 agenda has come to a successful conclusion, but the excitement continues. Tomorrow, the conference will continue to advance around the four major forums of AI models, AI infrastructure, AI applications, and embodied intelligence, and will welcome the highly anticipated PyTorch Day, with more heavyweight guests and first-line practical content coming soon, so stay tuned!